Sait Abdul Rahaman Wahab, AlBalawi Eid, Nagaraj Ramprasad
Department of Archives and Communication, King Faisal University, Hofuf, Kingdom of Saudi Arabia.
Department of Computer Science, College of Computer Science and Information Technology, King Faisal University, Al Hofuf, Kingdom of Saudi Arabia.
PLoS One. 2024 Dec 31;19(12):e0313386. doi: 10.1371/journal.pone.0313386. eCollection 2024.
Early Lung Cancer (LC) detection is essential for reducing the global mortality rate. The limitations of traditional diagnostic techniques cause challenges in identifying LC using medical imaging data. In this study, we aim to develop a robust LC detection model. Positron Emission Tomography / Computed Tomography (PET / CT) images are utilized to comprehend the metabolic and anatomical data, leading to optimal LC diagnosis. In order to extract multiple LC features, we enhance MobileNet V3 and LeViT models. The weighted sum feature fusion technique is used to generate unique LC features. The extracted features are classified using spline functions, including linear, cubic, and B-spline of Kolmogorov-Arnold Networks (KANs). We ensemble the outcomes using the soft-voting approach. The model is generalized using the Lung-PET-CT-DX dataset. Five-fold cross-validation is used to evaluate the model. The proposed LC detection model achieves an impressive accuracy of 99.0% with a minimal loss of 0.07. In addition, limited resources are required to classify PET / CT images. The high performance underscores the potential of the proposed LC detection model in providing valuable and optimal results. The study findings can significantly improve clinical practice by presenting sophisticated and interpretable outcomes. The proposed model can be enhanced by integrating advanced feature fusion techniques.
早期肺癌（LC）检测对于降低全球死亡率至关重要。传统诊断技术的局限性给利用医学影像数据识别LC带来了挑战。在本研究中，我们旨在开发一种强大的LC检测模型。利用正电子发射断层扫描/计算机断层扫描（PET/CT）图像来理解代谢和解剖数据，从而实现最佳的LC诊断。为了提取多个LC特征，我们对MobileNet V3和LeViT模型进行了改进。采用加权和特征融合技术来生成独特的LC特征。使用包括线性、三次和柯尔莫哥洛夫 - 阿诺德网络（KANs）的B样条在内的样条函数对提取的特征进行分类。我们使用软投票方法对结果进行集成。使用Lung-PET-CT-DX数据集对模型进行泛化。采用五折交叉验证来评估模型。所提出的LC检测模型实现了令人印象深刻的99.0%的准确率，损失最小为0.07。此外，对PET/CT图像进行分类所需资源有限。高性能突出了所提出的LC检测模型在提供有价值和最佳结果方面的潜力。研究结果通过呈现复杂且可解释的结果可显著改善临床实践。通过集成先进的特征融合技术可以增强所提出的模型。
